The　interpretation　of　natural　scenes,　generally　so　obvious　and　effortless　for　humans,　still　remains　a　challenge　in　computer　vision.　In　this　paper　we　propose　to　design　binary　classifiers　capable　to　recognize　some　generic　natural　scene　images,　the　countryside　class　and　the　city　class　for　instance.　After　the　segmentation　each　image　is　represented　by　a　set　of　regions.　Then　the　propounded　1*12　vectors　are　used　to　describe　the　regions.　Finally,　we　propose　a　new　simple　kernal　function　based　on　graph　edit　distance　and　raise　the　question　that　Munkres＇　algorithm　can　be　used　to　measure　the　similarity　between　the　images.　Experiments　show　that　the　improvements　are　effective.　©2010　IEEE.